Image classifier learning device, image classifier learning method, and program

Patent metadata

Field	Value
Publication number	US-11989929-B2
Application number	US-201917277248-A
Country	US
Kind code	B2
Filing date	Sep 6, 2019
Priority date	Sep 20, 2018
Publication date	May 21, 2024
Grant date	May 21, 2024

Abstract

An object is to make it possible to train an image recognizer by efficiently using training data that does not include label information. A determination unit 180 causes repeated execution of the followings. A feature representation model for extracting feature vectors of pixels is trained such that an objective function is minimized, the objective function being expressed as a function that includes a value that is based on a difference between a distance between feature vectors of pixels labeled with a positive example label and a distance between a feature vector of a pixel labeled with the positive example label and a feature vector of an unlabeled pixel, and a value that is based on a difference between a distance between a feature vector of a pixel labeled with the positive example label and a feature vector of an unlabeled pixel and a distance between a feature vector of a pixel labeled with the positive example label and a feature vector of a pixel labeled with a negative example label, and based on a distribution of feature vectors corresponding to the positive example label, a predetermined number of labels are given based on the likelihood that each unlabeled pixel is a positive example.

First claim

The invention claimed is:

1. An image recognizer training device comprising a processor configured to execute operations, comprising:
training a feature representation model based on a set of images that include an image including pixels each labeled with a positive example label or a negative example label and an image including unlabeled pixels, the feature representation model being a model for extracting feature vectors of the pixels and trained such that an objective function is minimized, the objective function being expressed as a function that includes a value that is based on a difference between a distance between feature vectors of pixels labeled with the positive example label and a distance between a feature vector of a pixel labeled with the positive example label and a feature vector of an unlabeled pixel, and a value that is based on a difference between a distance between a feature vector of a pixel labeled with the positive example label and a feature vector of an unlabeled pixel and a distance between a feature vector of a pixel labeled with the positive example label and a feature vector of a pixel labeled with the negative example label;
extracting feature vectors of a plurality of pixels included in the images based on the feature representation model trained;
estimating a distribution of feature vectors that are extracted with respect to pixels labeled with the positive example label;
calculating, with respect to each unlabeled pixel, a likelihood that the pixel is a positive example based on the distribution estimated and give the positive example label to pixels, a number of which is a first sample number determined in advance, in descending order of the likelihood and give the negative example label to pixels, the number of which is a second sample number determined in advance, in ascending order of the likelihood; and
determining whether or not there is an unlabeled pixel among pixels of the images, and if there is an unlabeled pixel, causes repeated execution of training, extraction of feature vectors, estimation, and labeling, wherein for each unlabeled pixel, the likelihood of the unlabeled pixel being a positive example is labeled based on a normal distribution and the feature vector of the unlabeled pixel.

2. The image recognizer training device according to claim 1 , wherein the objective function is expressed as a function that further includes a value that is based on a difference between a distance between feature vectors of pixels labeled with the same label and a distance between feature vectors of pixels labeled with different labels.

3. The image recognizer training device according to claim 1 , wherein the feature representation model includes a convolutional neural network.

4. The image recognizer training device according to claim 1 , wherein the objective function is associated with Triplet Loss.

5. An image recognizer training method comprising:
training a feature representation model based on a set of images that include an image including pixels each labeled with a positive example label or a negative example label and an image including unlabeled pixels, the feature representation model being a model for extracting feature vectors of the pixels and trained such that an objective function is minimized, the objective function being expressed as a function that includes a value that is based on a difference between a distance between feature vectors of pixels labeled with the positive example label and a distance between a feature vector of a pixel labeled with the positive example label and a feature vector of an unlabeled pixel, and a value that is based on a difference between a distance between a feature vector of a pixel labeled with the positive example label and a feature vector of an unlabeled pixel and a distance between a feature vector of a pixel labeled with the positive example label and a feature vector of a pixel labeled with the negative example label;
extracting feature vectors of a plurality of pixels included in the images based on the feature representation model trained;
estimating a distribution of feature vectors that arc extracted with respect to pixels labeled with the positive example label;
calculating with respect to each unlabeled pixel, a likelihood that the pixel is a positive example based on the distribution estimated and giving the positive example label to pixels, a number of which is a first sample number determined in advance, in descending order of the likelihood and giving the negative example label to pixels, the number of which is a second sample number determined in advance, in ascending order of the likelihood; and
determining whether or not there is an unlabeled pixel among pixels of the images, and if there is an unlabeled pixel, causing repeated execution of training, extraction of feature vectors, estimation, and labeling, wherein for each unlabeled pixel, the likelihood of the unlabeled pixel being positive example is labeled based on a normal distribution and the feature vector of the unlabeled pixel.

6. The image recognizer training method according to claim 5 , wherein the objective function is expressed as a function that further includes a value that is based on a difference between a distance between feature vectors of pixels labeled with the same label and a distance between feature vectors of pixels labeled with different labels.

7. The image recognizer training method according to claim 5 , wherein the feature representation model includes a convolutional neural network.

8. The image recognizer training method according to claim 5 , wherein the objective function is associated with Triplet Loss.

9. A computer-readable non-transitory recording medium storing computer-executable instructions that when executed by a processor cause a computer system to:
training a feature representation model based on a set of images that include an image including pixels each labeled with a positive example label or a negative example label and an image including unlabeled pixels, the feature representation model being a model for extracting feature vectors of the pixels and trained such that an objective function is minimized, the objective function being expressed as a function that includes a value that is based on a difference between a distance between feature vectors of pixels labeled with the positive example label and a distance between a feature vector of a pixel labeled with the positive example label and a feature vector of an unlabeled pixel, and a value that is based on a difference between a distance between a feature vector of a pixel labeled with the positive example label and a feature vector of an unlabeled pixel and a distance between a feature vector of a pixel labeled with the positive example label and a feature vector of a pixel labeled with the negative example label;
extracting extract feature vectors of a plurality of pixels included in the images based on the feature representation model trained by the representation learner;
estimating a distribution of feature vectors that are extracted with respect to pixels labeled with the positive example label;
determining, with respect to each unlabeled pixel, a likelihood that the pixel is a positive example based on the estimated distribution and give the positive example label to pixels, a number of which is a first sample number determined in advance, in descending order of the likelihood and give the negative example label to pixels, the number of which is a second sample number determined in advance, in ascending order of the likelihood; and
determining whether or not there is an unlabeled pixel among pixels of the images, and if there is an unlabeled pixel, causes repeated execution of

Assignees

• Nippon Telegraph & Telephone

Inventors

• Murasaki Kazuhiko
• Ando Shingo
• Sagata Atsushi

Classifications

• G06V10/7753 Primary

  Incorporation of unlabelled data, e.g. multiple instance learning [MIL] · CPC title

• G06F18/2155

  characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling · CPC title

• G06F18/22

  Matching criteria, e.g. proximity measures · CPC title

• G06V10/82

  using neural networks · CPC title

• G06T7/00

  Image analysis · CPC title